Research map / Architectures and efficiency / Neural architecture search
ScaleNet: Searching for the Model to Scale
Jiyang Xie, Xiu Su, Shan You, Zhanyu Ma, Fei Wang, Chen Qian
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add hierarchical sampling strategy (HSS) instead of uniform sampling (ablation) | ScaleNet · ImageNet-100 | top-1 accuracy (%): s0 84.16 -> 84.76; s4 90.02 -> 90.46 | Improved |
| add HSS instead of uniform sampling (ablation) | super-supernet training · path ranking | Pearson coefficient (%): 35.3 -> 73.6 | Improved |
| add HSS instead of uniform sampling (ablation) | super-supernet training · path ranking | Spearman and Kendall (%): 80.1 -> 83.9; 64.1 -> 66.2 | Improved |
| raise maximum scaling stage M from 1 to 3 (ablation) | ScaleNet MCEA · ImageNet-100 | top-1 accuracy (%): s0 84.06 -> 84.76; s4 90.30 -> 90.46 | Mixed |
| raise MCEA iterations T from 1 to 4 (ablation) | ScaleNet MCEA · ImageNet-100 | top-1 accuracy (%): s0 84.18 -> 84.76; s1 86.34 -> 87.18 | Improved |
| raise MCEA iterations T from 4 to 8 (ablation) | ScaleNet MCEA · ImageNet-100 | top-1 accuracy (%): s0 84.76 -> 84.50; s4 90.46 -> 90.36 | No clear effect |
| use proposed exponential generalization function instead of linear or squared (ablation) | ScaleNet s4 · ImageNet-100 | top-1 accuracy (%): 90.46 | Improved |
| search base model and scaling strategies jointly (ScaleNet) versus prior NAS scaling pipelines | ScaleNet · search cost | GPU/TPU days: 1,228 -> 485 | Improved |
| transfer ScaleNet-S0 instead of EfficientNet-B0 | fine-tuning · CIFAR-100 | top-1 accuracy (%): 88.1 -> 88.4 | Improved |
Closest papers by meaning
- A Hitchhiker's Guide to Scaling Law Estimation (ICML 2025 · similarity 0.85)
- Scaling Laws for Optimal Data Mixtures (NeurIPS 2025 · similarity 0.85)
- Search for Efficient Large Language Models (NeurIPS 2024 · similarity 0.85)
- Power Lines: Scaling laws for weight decay and batch size in LLM pre-training (NeurIPS 2025 · similarity 0.85)
- How to Scale Mixture-of-Experts: From muP to the Maximally Scale-Stable Parameterization (NeurIPS 2026 · similarity 0.84)
- Scaling Inference-Efficient Language Models (ICML 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
